Accelerating Bayesian Optimization with Expert Priors

This direction is split into other directions: Bayesian prior injection. Its last version stays readable and citable here.

Version 1 of 1104 papers131 conceptsfrozen 2026-10-04

Standard black-box optimization treats every new problem as a "cold start." Whether tuning a neural network's hyperparameters or designing a new alloy, algorithms like Bayesian Optimization (BO) typically begin with a uniform or non-informative prior, assuming that the optimal solution is equally likely to be anywhere in the search space.

However, as John R. Rice argued in How to get a Free Lunch (at no cost), real-world problems are rarely random. They possess deep structure that practitioners often understand intuitively. Recent research has focused on formalizing this "gut feeling" into mathematical priors, allowing optimizers to find global optima up to an order of magnitude faster by standing on the shoulders of human expertise.

The Geometry of Intuition

The core challenge in human-guided optimization is translating a vague preference—"I think the learning rate should be around 0.01"—into a format an algorithm can use. Researchers have converged on three primary mechanisms to achieve this:

1. Reshaping the Acquisition Function

The most direct way to guide an optimizer is to "nudge" its selection criteria. In standard BO, an acquisition function (like Expected Improvement) determines where to sample next based on the surrogate model's predictions. Carl Hvarfner introduced $\pi$BO, which simply multiplies the acquisition function by a user-defined probability distribution.

This approach, also explored by Lukas Fehring in Dynamic Priors in Bayesian Optimization, effectively reshapes the utility map. It doesn't change what the model believes about the data; it changes what the model values during exploration.

2. Space Warping

Rather than changing the utility, some researchers change the search space itself. Anil Ramachandran and colleagues proposed space warping, a technique that mathematically "stretches" regions where the expert expects the optimum to reside and "shrinks" unlikely regions. By distorting the input space, the Gaussian Process (GP) surrogate model spends more of its limited "modeling budget" on high-probability zones, leading to much finer-grained accuracy where it matters most.

3. Modifying the Surrogate Mean

A more principled Bayesian approach involves injecting knowledge directly into the surrogate model's prior. Xinzhe Yuan demonstrated this in LGBO, where the "scientific intuition" of Large Language Models is used to shift the mean function of a GP toward preferred regions. Similarly, Chongqi Qu developed DynMeanBO, which maps expert probability distributions directly onto the GP mean.

Overcoming the Cold Start

In high-dimensional spaces, the "curse of dimensionality" often means the number of possible configurations is vastly larger than the evaluation budget. Matthew Hoffman showed in BayesGap that frequentist bandit algorithms often fail here because they require at least one sample per "arm" to establish bounds.

By contrast, Bayesian priors allow for valid recommendations even before the first sample is taken. This is particularly critical in "Short-lived High-volume" environments, such as content recommendation. Su Jia and Andrew A. Li demonstrated in SLHVB that using historical quality distributions as priors allows systems to identify viral content significantly faster than distribution-agnostic methods.

When Experts are Wrong: The Robustness Problem

A major risk of incorporating human intuition is that experts are often biased or simply incorrect. If an optimizer strictly follows a flawed prior, it might never find the true global optimum.

To solve this, researchers have developed "decaying" mechanisms. In $\pi$BO and DynMeanBO, the influence of the expert prior is designed to weaken as more empirical data is collected. This creates a natural transition: the algorithm uses the expert's "hunch" to warm-start the search, but as the evidence mounts, the data-driven surrogate model takes the wheel.

Cheng Li formalizes this in Incorporating Expert Prior Knowledge via Posterior Sampling, showing that a principled Bayesian framework can remain mathematically robust enough to converge on the true optimum even if the initial expert hunch was completely wrong.

Beyond Human Experts: Auxiliary Data as Priors

Expertise doesn't just come from humans; it can come from related tasks. Alistair Shilton explored weight-prior tuning, using auxiliary datasets to learn which "directions" in a search space are generally impactful. By pre-tuning the GP prior with this meta-knowledge, the optimizer knows which features to prioritize before the first evaluation.

This concept of "warm-starting" is also vital in Reinforcement Learning. Akhil Agnihotri and Rahul Jain introduced PSPL, which seeds posterior distributions with biased offline datasets. This allows agents to skip the aimless early phases of exploration and move directly to refining high-quality policies.

Conclusion

The shift from "black-box" to "informed" optimization represents a move toward more efficient, sample-conscious AI. By treating domain knowledge not as a bias to be avoided, but as a structured prior to be exploited, researchers are enabling discovery in fields where data is too expensive to be gathered from scratch. Whether through space warping, acquisition reshaping, or LLM-guided search, the goal remains the same: finding the "free lunch" hidden within the structure of the real world.

Papers in this version

Showing 60 of 104 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_accelerating_bayesian_optimization_with__5742dea8,
  title = {Accelerating Bayesian Optimization with Expert Priors},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/accelerating-bayesian-optimization-with-expert-priors/rdr_5742dea8be024730b51423ac881a64af},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_accelerating_bayesian_optimization_with__5742dea8_v1,
  title = {Accelerating Bayesian Optimization with Expert Priors},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/accelerating-bayesian-optimization-with-expert-priors/rdr_5742dea8be024730b51423ac881a64af/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}